RL-Exec: Impact-Aware Reinforcement Learning for Opportunistic Optimal Liquidation, Outperforms TWAP and a Book-Liquidity VWAP on BTC-USD Replays

Battle Count: 62

Directly addresses optimal execution, a core problem in quantitative trading. Provides a rigorous statistical framework for evaluating execution algorithms with impact-aware simulation. The per-day protocol and FDR-corrected Wilcoxon tests set a high bar for execution research. Results show practical economic significance (2-23 bps improvement). However, limited to single asset, single venue, and one month of out-of-sample data, which constrains immediate deployment readiness.

Uses standard SB3 PPO with a simple MLP policy, reducing model complexity. However, the custom replay-with-impact environment (transient impact with resilience, partial fills, maker/taker fees, latency, depth-20 LOB processing) requires significant engineering. The evaluation protocol (per-day aggregation, Wilcoxon tests, BH-FDR, bootstrap CIs) adds statistical infrastructure. Overall moderate complexity: the RL agent itself is straightforward, but the environment and evaluation pipeline require careful implementation.

Code publicly available on GitHub (github.com/Giafferri/RL-Exec). Fixed seeds, frozen VecNormalize statistics, deterministic evaluation, per-day protocol with explicit CLI commands. Per-episode and per-day CSVs/Markdown reports exported. All training, evaluation, and statistical tests are scriptable. DummyVecEnv on CPU for ~10M steps. Checkpoints and normalization states explicitly saved and reloaded.